linux-kernel - Re: [PATCH v2 bpf-next 3/5] ftrace: introduce FTRACE_OPS_FL_SHARE

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20220715155953.4fb692e2@gandalf.local.home>
Date:   Fri, 15 Jul 2022 15:59:53 -0400
From:   Steven Rostedt <rostedt@...dmis.org>
To:     Song Liu <songliubraving@...com>
Cc:     Song Liu <song@...nel.org>, Networking <netdev@...r.kernel.org>,
        bpf <bpf@...r.kernel.org>, lkml <linux-kernel@...r.kernel.org>,
        Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andrii@...nel.org>,
        Kernel Team <Kernel-team@...com>,
        "jolsa@...nel.org" <jolsa@...nel.org>,
        "mhiramat@...nel.org" <mhiramat@...nel.org>
Subject: Re: [PATCH v2 bpf-next 3/5] ftrace: introduce
 FTRACE_OPS_FL_SHARE_IPMODIFY

On Fri, 15 Jul 2022 19:49:00 +0000
Song Liu <songliubraving@...com> wrote:

> > 
> > What about if we release the lock when doing the callback?  
> 
> We can probably unlock ftrace_lock here. But we may break locking order 
> with direct mutex (see below).

You're talking about the multi registering case, right?

> 
> > 
> > Then we just need to make sure things are the same after reacquiring the
> > lock, and if they are different, we release the lock again and do the
> > callback with the new update. Wash, rinse, repeat, until the state is the
> > same before and after the callback with locks acquired?  
> 
> Personally, I would like to avoid wash-rinse-repeat here.

But it's common to do. Keeps your hair cleaner that way ;-)

> 
> > 
> > This is a common way to handle callbacks that need to do something that
> > takes the lock held before doing a callback.
> > 
> > The reason I say this, is because the more we can keep the accounting
> > inside of ftrace the better.
> > 
> > Wouldn't this need to be done anyway if BPF was first and live kernel
> > patching needed the update? An -EAGAIN would not suffice.  
> 
> prepare_direct_functions_for_ipmodify handles BPF-first-livepatch-later
> case. The benefit of prepare_direct_functions_for_ipmodify() is that it 
> holds direct_mutex before ftrace_lock, and keeps holding it if necessary. 
> This is enough to make sure we don't need the wash-rinse-repeat. 
> 
> OTOH, if we wait until __ftrace_hash_update_ipmodify(), we already hold
> ftrace_lock, but not direct_mutex. To make changes to bpf trampoline, we
> have to unlock ftrace_lock and lock direct_mutex to avoid deadlock. 
> However, this means we will need the wash-rinse-repeat. 
> 
> 
> For livepatch-first-BPF-later case, we can probably handle this in 
> __ftrace_hash_update_ipmodify(), since we hold both direct_mutex and 
> ftrace_lock. We can unlock ftrace_lock and update the BPF trampoline. 
> It is safe against changes to direct ops, because we are still holding 
> direct_mutex. But, is this safe against another IPMODIFY ops? I am not 
> sure yet... Also, this is pretty weird because, we are updating a 
> direct trampoline before we finish registering it for the first time. 
> IOW, we are calling modify_ftrace_direct_multi_nolock for the same 
> trampoline before register_ftrace_direct_multi() returns.
> 
> The approach in v2 propagates the -EAGAIN to BPF side, so these are two
> independent calls of register_ftrace_direct_multi(). This does require
> some protocol between ftrace core and its user, but I still think this 
> is a cleaner approach. 

The issue I have with this approach is it couples BPF and ftrace a bit too
much.

But there is a way with my approach you can still do your approach. That
is, have ops_func() return zero if everything is fine, and otherwise returns
a negative value. Then have the register function fail and return whatever
value that gets returned by the ops_func()

Then have the bpf ops_func() check (does this direct caller handle
IPMODIFY? if yes, return 0, else return -EAGAIN). Then the registering of
ftrace fails with your -EAGAIN, and then you can change the direct
trampoline to handle IPMODIFY and try again. This time when ops_func() is
called, it sees that the direct trampoline can handle the IPMODIFY and
returns 0.

Basically, it's a way to still implement my suggestion, but let BPF decide
to use -EAGAIN to try again. And then BPF and ftrace don't need to have
these special flags to change the behavior of each other.

-- Steve